Harnessing Interpretable Deep Learning to Predict Meropenem Resistance in Klebsiella pneumoniae

Antimicrobial resistance represents an escalating global healthcare threat, complicating treatment and increasing both morbidity and mortality. Deep learning offers promising solutions, particularly for bacterial profiling using omics data. For instance, classifying bacterial strains as resistant or susceptible to antibiotics depends on identifying genomic signatures associated with resistance mechanisms. This study introduces DeepMDC, a deep learning architecture for bacterial profiling that utilizes whole-genome data as input. In genomics, obtaining precise labels for each gene or mutation is costly and often ambiguous; therefore, labels are typically aggregated. Consequently, phenotypic classification is framed as a Multiple-instance learning problem, where training data comprise sets (bags) of instances with a single label per bag. The core component of DeepMDC is a Modern Hopfield Network, and its input encompasses all open reading frames (ORFs) identifiable in a genome, including small ORFs. A key feature of DeepMDC is its interpretability, which facilitates model validation and hypothesis generation regarding bacterial resistance by leveraging the attention mechanism inherent to Modern Hopfield Networks. Evaluation is centered on Klebsiella pneumoniae and meropenem due to their clinical significance. DeepMDC achieves values exceeding 0.9 for several performance metrics, including ROC AUC, balanced accuracy, and F1 score. Several well-known resistance-associated genes, such as the Tn3-like element Tn4401 family transposase, received high attention scores during the inference process. The attention mechanism also suggests a potential role for small ORFs in meropenem resistance. Results are discussed in detail, including analysis of false positives and negatives.